Supervised Fine-tuning with Synthetic Rationale Data Hurts Real-World Disease Prediction

The study published in arXiv examines the impact of supervised fine-tuning (SFT) with synthetic rationale data on Alzheimer's disease prediction, revealing that this approach consistently degrades model performance compared to label-only fine-tuning across various configurations and model families. Despite the generated rationales being medically accurate, the research identifies a structural conflict between narrative plausibility and discriminative optimization as the root cause of this performance decline. This work highlights the need for a nuanced understanding of rationale-based supervision in clinical applications, informing practitioners on when such techniques may be counterproductive.